Intelligent control of nonlinear dynamical systems using multilayer neural networks

This dissertation focuses on two major topics: control of a class of nonlinear systems with known dynamics and the control of systems with unknown dynamics.; Specifically in the first part of this work, a systematic approach for modeling and base motion control of a mobile base with an onboard robot arm is presented. Feedback linearization is used to take into account the complete dynamics with nonholonomic constraints, yet methods from potential field theory are incorporated to provide resolution among possibly conflicting performance goals (e.g. path following and obstacle avoidance). The feedback linearization provides an inner loop that accounts for possible motion of the onboard arm. A rigorous yet simple approach to motion planning through optimization techniques is then presented for these mobile vehicles. The resulting Cartesian trajectory generated from the motion planning algorithm is employed as the reference trajectory in the outer loop, which is designed based on a Lyapunov function candidate. The net result is a base motion controller that enables these mobile vehicles to track a Cartesian trajectory with a desired final orientation (docking angle) and provides the necessary intelligence to avoid any static obstacles by using a simple correction term generated from a potential function.; In the second part of this work a family of novel learning algorithms is proposed for the control of multi-input and multi-output (MIMO) unknown nonlinear dynamical systems by employing neural networks (NN). The structure of the NN controller is derived using a filtered error/passivity approach. For guaranteed stability, it is shown that the learning rate parameter in the case of the delta rule should decrease with the number of hidden-layer neurons. The notion of persistency of excitation (PE) for a multilayer NN is defined and explored. New on-line weight tuning algorithms are developed that do not need the persistency of excitation condition. The notions of discrete-time "passive NN" and "robust NN" are introduced. These multilayer NN weight tuning algorithms are extended for the Model Reference Adaptive Control (MRAC) of a class of nonlinear systems. Finally, the identification using NN of a class of nonlinear dynamical systems is discussed.